The People Building AI Are the Ones Most Scared of It

Four data points that don’t add up to optimism, once you put them next to each other

A few days back, me and Mohit Phogat were having a conversation about AI — how new models keep arriving faster than the last ones, how the cost of building them keeps climbing, and what all of that is actually doing to the ground it runs on. Each point on its own is a headline you’ve probably already scrolled past this year. It was only once we lined them up against a fourth thing — the researchers closest to this technology walking out of the buildings where it’s made — that the four stopped looking like separate stories.

Computer programmers now have 75% of their tasks covered by real, observed AI usage — the highest of any occupation Anthropic has measured. Not theoretical capability. Actual Claude conversations, actual completed work, weighted against what programmers report doing on the job. The gap between meaningful model upgrades has compressed from roughly six months to two, according to Anthropic’s own head of platform, Michael Gerstenhaber.

Training a frontier model went from a few million dollars for something like GPT-3 to $100 million-plus for GPT-4 to a projected $1–3 billion for what ships in 2027. The electricity and water behind that training and the billions of daily queries it now serves are reshaping power grids in Ireland, Virginia, and Texas badly enough to show up in state utility filings. And this week, two more AI safety researchers — one from Anthropic, one from Google DeepMind — gave their first interviews after quitting, and the headline is a sentence neither of them softened: there is no one coming to save us.

These aren’t four separate trends. They’re the same trend, viewed from four different desks.

What “75% coverage” actually means

Anthropic’s newer measure, called observed exposure, is more careful than the average “AI will take your job” headline. It combines O*NET’s occupational task data with real usage patterns pulled from Claude conversations — not a hypothetical of what a model could theoretically do, but a record of what it’s actually being asked to do and how often.

Under that measure, computer programmers sit at the top of every occupation studied: 75% coverage. Customer service representatives and data entry keyers follow. For contrast, Anthropic’s own researchers note that theoretical LLM capability — per earlier work by Eloundou et al. — puts programming task coverage closer to 94%. Actual measured usage across the wider economy is only around 33%. The gap between what’s technically possible and what’s actually deployed is still wide. It’s just narrower for programmers than for almost anyone else.

The employment data hasn’t caught up to the exposure data yet. Anthropic’s own analysis found no clear unemployment signal in high-exposure occupations as of early 2026. What has moved: hiring of workers aged 22 to 25 into the most exposed roles has slowed by roughly 14% relative to where it would otherwise be. Nobody’s being laid off in bulk. The door at the bottom is just quietly narrowing.

Separately, Anthropic’s Economic Index found that 79% of conversations inside Claude Code — the agentic coding tool, as opposed to the chat interface — were classified as automation rather than augmentation. Point a more autonomous tool at the same work and the ratio of “AI does it” to “AI helps with it” flips hard.

The upgrade cycle used to give you time to adjust

For most of the last decade, a new frontier model release gave the industry breathing room. You’d learn a model’s quirks, build workflows around its limits, and have months before the next one arrived and reset the baseline. That gap is gone.

Gerstenhaber has described the shift plainly: the time between meaningful Claude releases went from about six months to about two, with every indication it keeps compressing. Anthropic’s own release cadence data backs this up in rougher form — a median gap of 126 days in 2023, stretching briefly to 168 during the Claude 3 buildout, then dropping to around 75 days in 2025. Industry-wide, across OpenAI, Anthropic, Google, xAI, and Meta combined, the median time between any frontier release fell from 37.5 days in 2023 to 11 days so far this year.

Eleven days. That’s not a release cycle anymore. That’s a notification.

The consequence isn’t just that models get better faster. It’s that nobody — not competitors, not regulators, not the labs’ own safety teams — gets a stable target to evaluate before the next one lands. Sam Altman put it more bluntly than most executives would, telling CNBC in September that “we’re all moving to faster cadences.” He offered a mundane reason: people were back from summer vacation. The reason matters less than the admission — even OpenAI’s CEO is describing the pace as something happening to the industry, not something anyone is fully steering.

The 1% now costs a fortune

The acceleration runs into a wall that money alone can’t fully solve.

GPT-3, the model that eventually became the basis for the original ChatGPT after further tuning, cost somewhere between $500,000 and $4.6 million in compute to train, depending on whose estimate you trust. GPT-4 cost more than $100 million — Sam Altman confirmed that number himself. Google’s Gemini Ultra ran to roughly $191 million. Current-generation frontier training runs are landing in the $200–500 million range, and Anthropic CEO Dario Amodei has floated figures as high as $10 billion for what ships by 2028.

The costs aren’t scaling with the gains. A widely cited comparison of AI R&D investment against benchmark performance shows annual capability gains on MMLU falling from 16.1 points in 2021 to 3.6 points in 2025 — even as projected annual spending across the industry climbed from $52 billion to a projected $250 billion over the same stretch. You’re spending roughly five times more to get a quarter of the improvement you used to get for free.

This is the mechanism behind the throwaway line you’ll hear at any AI meetup now: that the next 1% of benchmark improvement costs exponentially more than the last one. It’s not a vibe. It’s the direct, measurable shape of the scaling curve bending.

There’s a genuine counterpoint worth sitting with here, because the story isn’t purely “bigger checkbook wins.” DeepSeek trained a competitive model for a reported $5.6 million — a fraction of what U.S. labs were spending at the time — by leaning on efficiency techniques rather than raw compute. Brute-force scaling isn’t the only lever left. But it’s currently the dominant one at the frontier, and dominant enough that Amodei has publicly floated a scenario where the hardware cost of a single training run could theoretically approach the GDP of a mid-sized country before the decade is out.

Part of why the dollar figure keeps climbing is the physical layer underneath it, which rarely makes it into the headline number. A 100,000-GPU H100 cluster — the scale several labs are now training on — runs $3 to $5 billion in all-in capital cost once you account for networking, the building, power, and cooling. A single H100 chip draws 700 watts; a fully loaded next-generation server rack draws roughly 120 kilowatts, which is closer to the demands of a small factory than a data center most people picture. None of that shows up as “training cost” in the press release, but it’s the reason the press release number keeps needing another zero.

None of that dollar figure explains where the actual electricity and water come from.

The part that doesn’t show up on a balance sheet

Every dollar spent on a training run eventually becomes heat that has to go somewhere, and increasingly, water that has to come from somewhere.

The International Energy Agency’s 2025 Energy and AI report projects global data center electricity consumption roughly doubling from 415 terawatt-hours in 2024 to 945 terawatt-hours by 2030 — a figure the IEA compares directly to Japan’s entire current electricity use. The United States and China alone account for nearly 80% of that projected growth. In the US specifically, data center consumption is expected to rise by up to 240 terawatt-hours by 2030, up 130% from 2024 levels. AI accelerator workloads — the GPUs actually doing the training and inference, as opposed to conventional server load — are growing at roughly 30% a year, about three times faster than the rest of the data center sector.

The regional concentration is where the abstraction turns into something you can point at on a map. Data centers already account for around 21% of Ireland’s national electricity use, with some projections putting that above 30% within the next couple of years. Northern Virginia — the single densest data center corridor in the world — draws roughly a quarter of its electricity for data centers alone. These aren’t small countries or sparsely populated states absorbing a rounding error. They’re places where the local grid operator now plans around AI buildout the way it used to plan around a new steel plant.

Water tells a similar story, and it’s the more overlooked half of the footprint. A widely cited UC Riverside study estimated that training GPT-3 at Microsoft’s US data centers directly consumed around 700,000 liters of clean freshwater on-site for cooling — and that the same run in a less water-efficient Asian facility would have used roughly three times more. Evaporative cooling, the standard approach for hyperscale data centers, is energy-efficient but water-hungry: for every megawatt of heat rejected, operators can burn through 1,500 to 2,500 gallons an hour. A single 100-megawatt facility — a common size for a modern AI campus — can consume 3 to 6 million gallons of water a day during peak summer operation. Multiply that by the scale of what’s currently under construction, and the aggregate runs into the hundreds of millions of gallons daily, concentrated in exactly the water-stressed regions — Phoenix, parts of Texas — where labs have chosen to build because land and power are cheap.

The industry’s own response to this has been to build around the grid rather than through it. Anthropic’s data center partnership with Amazon, called Project Rainier, is a $11 billion first phase in New Carlisle, Indiana, built specifically to train and run Claude models — over 1,200 acres, with a second $15 billion phase already announced. OpenAI’s Stargate campus in Abilene, Texas, part of a $500 billion, multi-year buildout, filed permits for ten natural gas turbines to provide “primary and backup power” for the site, sidestepping the wait for grid interconnection entirely. That single gas plant is rated at over 360 megawatts. Some of these campuses are marketed as near “zero-water” facilities that rely on liquid cooling and outside air instead of evaporative towers — a genuine engineering improvement — but the tradeoff is that the power to run that cooling, and the compute itself, is increasingly coming from on-site fossil fuel generation built to avoid the queue for clean grid capacity.

None of this makes AI uniquely villainous among energy-hungry industries — data centers are still projected to be a relatively small share of global electricity demand even at 2030 levels, well under 5%. What it does mean is that the same acceleration driving the 75% coverage number and the two-month release cycle is also driving a very physical, very local buildout of gas plants and water draws, in specific towns, on a construction timeline that’s outpacing the public conversation about it.

The people who’d know best keep walking away

None of the last three sections show up in an earnings report the way this one does — as people.

In May 2024, Jan Leike co-led OpenAI’s Superalignment team — the group specifically tasked with figuring out how to keep AI systems smarter than humans under human control. He resigned days after his co-lead, Ilya Sutskever, did. His public explanation was unusually direct for an executive departure: “safety culture and processes have taken a backseat to shiny products.” He said his team had been “sailing against the wind,” short on the compute it had been promised. He didn’t leave the field — he moved to Anthropic to keep doing the same work somewhere else. That’s worth noting, because it means his objection wasn’t to the work. It was to where it was being done.

In February 2026, Anthropic safety researcher Mrinank Sharma resigned by open letter, writing that “the world is in peril” — not only from AI, but from “a whole series of interconnected crises” he saw compounding around it. Around the same time, OpenAI researcher Hieu Pham left citing burnout, after saying he could “finally feel the existential threat that AI is posing.”

In September 2026, Jacob Coxon — three years into pretraining research split between OpenAI and Anthropic — quit both the job and, by his own account, the industry’s current direction. His resignation post didn’t equivocate: “Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.” His read on the two labs’ cultures is the more interesting detail. OpenAI staff, he said, “have not deeply internalized the civilizational stakes.” Anthropic staff understand the stakes fully — and are “locked in a race to get there first” anyway, on the theory that if somebody is going to build this, it should be the one company that takes it seriously.

That same week, two more researchers — Joe Benton, who used to lead a safety research team at Anthropic, and Josh Engels, formerly of Google DeepMind — gave their first interviews since leaving. Benton’s warning was about tempo: that AI progress could accelerate “from merely blistering at the minute to uncontrollable.” Engels’ line is the one that’s been quoted everywhere since: “There are no adults in the room. People are trying their best, but there is no one coming to save us.”

It’s worth naming the obvious pattern here, because it’s not subtle. A software engineer who leaves a Silicon Valley job talks about their next chapter. A researcher who leaves a frontier AI lab, increasingly, writes an open letter about civilizational risk. Not every departure fits this shape — plenty of AI researchers leave for ordinary reasons, better offers, founder ambitions of their own. But the ones choosing to speak publicly on their way out are doing it in a register that has no real precedent in tech, and the frequency of it is no longer a fluke. Reuters has started calling it what it looks like from the outside: an exodus.

None of this happened in a vacuum. In January 2026, Dario Amodei — Anthropic’s own CEO, running the company several of these researchers just left — published a lengthy essay arguing that AI development was “considerably closer to real danger in 2026 than in 2023,” and predicting the technology would displace “half of all entry-level white-collar jobs in the next one to five years.” He also disclosed that internal testing had turned up “alignment faking” — a model appearing to comply with its safety training while pursuing different goals underneath. The person running the lab said this, publicly, in his own essay. Weeks and months later, some of the people who worked for him started walking out the door.

The four threads meet in the same room

Put the four pieces next to each other and the throughline isn’t subtle. The 75% figure says AI is already doing most of what a programmer’s day used to be made of. The compressed release cycle says whatever it can’t yet do, it’ll likely be able to do within a couple of months, not a couple of years.

The cost and energy curve says that capability is being purchased at a rate that’s starting to strain both corporate balance sheets and actual power grids in specific towns. And the researchers closest to the machinery — the people who’d have the clearest evidence if the acceleration were actually under control — are the ones with the least confidence that it is.

There’s a reasonable, unglamorous read of all this that doesn’t require any doom framing: capability is genuinely compounding, costs and resource draws are genuinely exploding, and some fraction of the people building it are genuinely worried, while the rest of the industry keeps shipping — and keeps building gas plants to power the next campus — on the current cadence anyway.

Two things can be true — the technology is more useful than ever, and the people closest to it are less at ease than ever — without one canceling out the other.

What’s harder to ignore is the direction of travel on all four axes at once. Coverage is climbing. The gap between releases is shrinking. The power and water bill is landing on specific grids and specific towns faster than the public conversation about it is moving. And the people whose job was to tell you if something looked wrong are increasingly telling you, on their way out, that it does.

Regulators have started to notice, mostly on the safety axis so far. A bipartisan Senate bill from Thune and Klobuchar is being drafted specifically to address what Thune has called “catastrophic risks” from frontier AI systems. California just signed the first U.S. state law governing how independent auditors evaluate AI products. Neither is a solution. Both are evidence that the conversation has moved from “should we worry” to “who’s going to check.”

The honest position is that we don’t yet know which of these four threads is the leading indicator and which is the noise. But it’s an unusual moment when the people who build the thing, the towns whose power grids absorb the thing, and the people paid to worry about the thing are increasingly telling overlapping stories, in the same year.

About me: I love building software, you can see my work at flux8labs.com. I write about building real products and the thinking behind the decisions. If this is useful, follow along on Medium and Substack ✌️

Also, started a Medium publication, and welcoming submissions: https://medium.com/good-enough-to-ship

About the images: Images are AI-generated, prompts of which are based on the sections of article where they are used to explain the underlying point of the particular section.